Tire surface defect detection method and system based on multispectral image fusion

Abstract

This invention discloses a method and system for tire surface defect detection based on multispectral image fusion, belonging to the field of image recognition technology. It includes: loading the tire model identifier of the tire to be detected; retrieving the prior structure map corresponding to the tire model identifier from a tire model prior structure map database; obtaining a real-time multispectral image; performing non-rigid registration between the prior structure map and the real-time multispectral image to obtain an aligned prior map; dynamically dividing the real-time multispectral image into regions of interest (ROIs) and assigning a corresponding detection strategy to each ROI; performing detection on the corresponding ROI using the detection strategy for each ROI to obtain the detection result for each ROI; summarizing the detection results of all ROIs, and making a final judgment on the summarized detection results based on the historical distribution information of defects in the aligned prior map, outputting the tire quality assessment result. This invention effectively improves the accuracy of tire surface defect detection.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically to a method and system for detecting tire surface defects based on multispectral image fusion. Background Technology

[0002] Tire surface defect detection is a crucial step in ensuring tire quality. Currently, multispectral imaging-based detection technologies are being gradually applied. Existing technologies, such as CN119246521A, use a movable probe to acquire multispectral images along the interior of the product, using depth location information to pinpoint internal defects. Another example is CN119164955A, which fuses multispectral images with 3D point clouds to construct a high-precision 3D model and combines this with path optimization algorithms to achieve defect detection.

[0003] However, the above methods have shortcomings in tire inspection scenarios. For example, the tire surface has complex backgrounds such as characters and patterns, and existing methods are prone to misjudging the edges of characters as cracks, resulting in a high false alarm rate. Different tire models have large structural differences, making it difficult for the detection system to quickly adapt and switch. Furthermore, using a uniform detection strategy for different areas such as character areas and the bottom of grooves results in wasted computing power and insufficient detection accuracy in key areas. Summary of the Invention

[0004] This invention provides a method and system for detecting tire surface defects based on multispectral image fusion, aiming to solve the technical problem of insufficient accuracy in tire surface defect detection in the prior art.

[0005] In view of the above problems, the present invention provides a method and system for detecting tire surface defects based on multispectral image fusion.

[0006] In a first aspect, the present invention provides a tire surface defect detection method based on multispectral image fusion, comprising: Load the tire model identification of the tire to be tested; Based on the tire model identifier, the prior structure map corresponding to the tire model identifier is retrieved from the pre-built tire model prior structure map database, wherein the prior structure map includes geometric structure information, optical prior information and defect history distribution information; Acquire multispectral images of the tire to be tested to obtain real-time multispectral images; The prior structure map is non-rigidly registered with the real-time multispectral image to obtain an aligned prior map that is aligned with the real-time multispectral image. Based on the aligned prior map, the real-time multispectral image is dynamically divided into several regions of interest, and a corresponding detection strategy is assigned to each region of interest. The detection strategies include a high-precision defect detection strategy, an optical character recognition strategy, a rapid screening strategy, and a hidden area enhancement detection strategy. Each region of interest is detected using the detection strategy corresponding to that region, and the detection results for each region of interest are obtained. The detection results of all regions of interest are summarized, and a final judgment is made on the summarized detection results based on the historical distribution information of defects in the aligned prior map, and the tire quality judgment result is output.

[0007] Secondly, the present invention provides a tire surface defect detection system based on multispectral image fusion, comprising: The identifier acquisition module is used to load the tire model identifier of the tire to be detected; The map retrieval module is used to retrieve the prior structure map corresponding to the tire model identifier from a pre-built tire model prior structure map database based on the tire model identifier. The prior structure map includes geometric structure information, optical prior information, and defect history distribution information. The image acquisition module is used to acquire multispectral images of the tire to be inspected, and obtain real-time multispectral images; The map alignment module is used to perform non-rigid registration between the prior structure map and the real-time multispectral image to obtain an aligned prior map that is aligned with the real-time multispectral image. The region segmentation module is used to dynamically divide the real-time multispectral image into several regions of interest based on the aligned prior map, and assign a corresponding detection strategy to each region of interest. The detection strategies include a high-precision defect detection strategy, an optical character recognition strategy, a rapid screening strategy, and a hidden area enhancement detection strategy. The region detection module is used to detect the corresponding region of interest by using the detection strategy corresponding to each region of interest, and to obtain the detection result of each region of interest; The result output module is used to summarize the detection results of all regions of interest, make a final judgment on the summarized detection results based on the defect history distribution information in the aligned prior map, and output the tire quality judgment result.

[0008] One or more technical solutions provided in this invention have at least the following technical effects or advantages: This invention provides a tire surface defect detection method and system based on multispectral image fusion. By loading the tire model identifier and matching and calling the corresponding prior structure map, and relying on preset geometric structure, optical priors, and historical defect distribution information, a precise prior benchmark is established for detection. Then, real-time multispectral images are acquired and non-rigid registration is performed, which can effectively eliminate detection deviations caused by tire position offset and shape deformation, and achieve precise alignment between the prior map and the real-time image. Furthermore, based on the aligned prior map, regions of interest are dynamically divided and differentiated detection strategies are assigned. High-precision detection, character recognition, rapid screening, and hidden area enhancement modes can be adapted to different regional characteristics, balancing detection efficiency and targeting. By executing corresponding detection strategies in different regions, the defect characteristics of each region can be accurately captured, comprehensively improving the detection capability of different types of defects. Finally, the detection results are summarized and combined with the historical defect distribution information to complete the final judgment, which can effectively filter false detections and enhance the accuracy of judgment, ultimately achieving intelligent detection of tire surface defects with high efficiency, accuracy, and low false alarms. Attached Figure Description

[0009] Figure 1 A schematic flowchart of a tire surface defect detection method based on multispectral image fusion provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a tire surface defect detection system based on multispectral image fusion provided in an embodiment of the present invention; The components represented by each number in the attached diagram are explained below: The module includes: identifier acquisition module 11, map retrieval module 12, image acquisition module 13, map alignment module 14, region division module 15, region detection module 16, and result output module 17. Detailed Implementation

[0010] This invention provides a method and system for detecting tire surface defects based on multispectral image fusion, which addresses the technical problem of insufficient accuracy in tire surface defect detection in existing technologies.

[0011] Example 1, as Figure 1 As shown, this invention provides a tire surface defect detection method based on multispectral image fusion, the method comprising: S100: Load the tire model identifier of the tire to be tested.

[0012] In this embodiment of the invention, a tire model identifier of the tire to be tested is loaded. The tire model identifier is obtained by reading the barcode or QR code on the surface of the tire to be tested using a barcode reader. Different tire models exhibit significant differences in tread patterns, sidewall structures, dimensions, optical reflection characteristics, and defect-prone areas. If the specific model of the tire to be tested cannot be accurately determined, the corresponding prior structure map cannot be matched, leading to a lack of detection benchmarks, incorrect area division, and failure of detection strategy adaptation, ultimately reducing the accuracy and applicability of defect detection. Therefore, it is necessary to first load a unique model identifier of the tire to be tested in a standardized manner to provide a matching basis for the subsequent entire testing process.

[0013] Specifically, the barcode reader is pre-installed at a designated location in the tire inspection station, ensuring that the reader's image capture angle is aligned with the fixed area on the tire surface where the code is printed or pasted. After the tire to be inspected is transported to the inspection station, the barcode reader is activated to capture images of the barcode or QR code on the tire surface. The internal parsing module of the barcode reader decodes the captured code and extracts the corresponding tire model information from the code. The inspection system receives the parsing results transmitted by the barcode reader and completes the loading and storage of the tire model identification.

[0014] For example, taking a 215 / 60R17 passenger car tire produced by a certain manufacturer as an example, the tire sidewall has a corresponding QR code printed at a fixed position, and the tire model identifier bound to the code is LT-21560R17-M02; after the tire to be inspected is transported to the inspection station, the industrial barcode reader at the station is aligned with the QR code area on the tire sidewall to complete the image acquisition. After the barcode reader parses the QR code, it outputs the model identifier LT-21560R17-M02. The inspection system successfully loads this identifier as the matching basis for subsequent processes.

[0015] In this embodiment of the invention, the tire model identifier is obtained by automatically reading the code through a barcode reader, replacing manual input and avoiding human input errors; it can quickly and accurately obtain the unique identification information of the tire to be tested, ensuring the targeting and standardization of the testing process, and improving the efficiency and accuracy of the pre-test preparation.

[0016] S200: Based on the tire model identifier, retrieve the prior structure map corresponding to the tire model identifier from the pre-built tire model prior structure map database, wherein the prior structure map includes geometric structure information, optical prior information and defect history distribution information.

[0017] In this embodiment of the invention, based on the tire model identifier, a priori structure map corresponding to the tire model identifier is retrieved from a pre-constructed priori structure map database. The priori structure map includes geometric structure information, optical prior information, and historical defect distribution information. Different tire models exhibit significant differences in surface geometry, optical reflection characteristics, and the location and distribution patterns of defects. Direct defect detection without a unified benchmark can easily lead to problems such as chaotic region division, incorrect detection strategy adaptation, omission of hidden defects, and misjudgment of normal textures. By pre-constructing a priori structure map database containing geometric structure information, optical prior information, and historical defect distribution information, and accurately retrieving the corresponding map based on the tire model identifier, a standardized priori benchmark can be provided for subsequent image registration, region division, strategy allocation, and defect judgment, ensuring the targetedness, accuracy, and reliability of the detection from the source.

[0018] Step S200 in the method provided in this embodiment of the invention includes: The construction of a priori structure map database for tire models includes: For each tire model, the geometric structure information of the tire model is extracted, wherein the geometric structure information includes the tread pattern outline, the coordinates of the sidewall character area, the groove depth distribution map, and the tread rubber thickness distribution map; Images of standard tire models of the specified tire type were collected under multiple spectra, and typical reflectance ranges and texture feature vectors of each region under different spectra were extracted as optical prior information. Summarize the historical defect detection data of the tire models, mark the actual areas where defects occurred, and generate a defect probability heatmap as historical defect distribution information; The geometric structure information, the optical prior information, and the defect history distribution information are associated and stored with the tire model to construct a tire model prior structure map database.

[0019] First, for each tire model, the geometric structure information of that tire model is extracted. This geometric structure information includes the tread pattern outline, the coordinates of the sidewall character area, the groove depth distribution map, and the tread rubber thickness distribution map. The tread pattern outline refers to the boundary of the texture pattern in the area where the tire surface contacts the ground, used to distinguish between tread protrusions and grooves. The sidewall character area coordinates refer to the precise location information of characters such as specifications and brand printed on the tire sidewall, used to avoid textual interference. The groove depth distribution map refers to the depth data of the tread grooves at different locations, used to identify hidden areas with insufficient light. The tread rubber thickness distribution map refers to the spatial distribution of the tread rubber layer thickness, used to assess the degree of wear and aging risk.

[0020] Specifically, for each tire model, its geometric structure information is extracted using the following methods: First, obtain the CAD design drawing of the tire model or scan a standard tire using a high-precision 3D scanner to generate 3D point cloud data of the tire surface. Based on the point cloud data, use curvature analysis and region growing algorithms to extract the raised areas and recessed grooves of the tread pattern, forming a binary boundary map of the tread pattern outline.

[0021] Secondly, in the tire sidewall area, morphological text region detection or deep learning-based text detection models are used to identify the bounding box coordinates of character blocks, record the coordinates of the smallest bounding rectangle vertex of each character region, and form a set of tire sidewall character region coordinates.

[0022] Next, regarding the groove depth, the vertical distance from the bottom of the groove to the top of the adjacent tread block in the 3D point cloud data is sampled and interpolated to generate a groove depth distribution map with the same resolution as the tread unfolded map, with each pixel position corresponding to a depth value.

[0023] Finally, for the tread rubber thickness, a gridded measurement is performed along the tire's circumference and lateral direction using an ultrasonic thickness gauge or optical coherence tomography (OCT) device, and the measurement results are interpolated to generate a tread rubber thickness distribution map.

[0024] For example, for a tire with the model number LT-21560R17-M02, the following geometric structure information is extracted: tread pattern outline: boundary coordinates of the four longitudinal main grooves and the lateral auxiliary grooves; sidewall character area coordinates: pixel coordinates of the upper left corner (25,180) and lower right corner (180,320) of the character area; groove depth distribution map: global distribution data of main groove depth 8mm and auxiliary groove depth 5mm; tread rubber thickness distribution map: distribution data of rubber thickness of 10mm at the tread protrusions and rubber thickness of 7mm at the groove edges.

[0025] Secondly, images of standard tires of the specified model were acquired under multiple spectra. Typical reflectance ranges and texture feature vectors for each region under different spectra were extracted as optical prior information. Multiple spectra refer to illumination acquisition spectra including different bands such as visible light and near-infrared. Typical reflectance range refers to the stable reflectance value range of different regions of the tire under the corresponding spectrum. Texture feature vectors are feature arrays composed of texture parameters such as contrast, roughness, and directionality.

[0026] Specifically, for each tire model, images of a standard tire of that model are collected under multiple spectra. The specific steps are as follows: In a controlled lighting environment, a multispectral imaging system is used to photograph the standard tire in visible light, near-infrared, and short-wave infrared bands, ensuring coverage of key areas such as the tread, shoulder, and sidewall. Geometric correction and registration are performed on the images under each spectrum to align the pixels of different spectral images. Guided by the extracted geometric structure information, the images are divided into different categories such as tread block regions, groove regions, and character regions. For each category, the pixel grayscale value distribution under different spectra is statistically analyzed, and the mean and standard deviation of reflectance under each spectrum are calculated to obtain the typical reflectance range. At the same time, the texture features of each category are extracted using a gray-level co-occurrence matrix or Gabor filter bank, and indices such as contrast, roughness, and directionality are calculated to form a texture feature vector. The reflectance ranges and texture feature vectors of all categories under all spectra are structured and stored as optical prior information for that model.

[0027] For example, visible light and near-infrared images were collected from the LT-21560R17-M02 standard tire: tread pattern area: visible light reflectance 30%~40%, near-infrared reflectance 45%~55%; sidewall area: visible light reflectance 25%~35%; texture feature vector: feature array with tread area contrast 85, roughness 12, and directionality 0.75.

[0028] Next, the historical defect detection data for the aforementioned tire models is compiled, the actual defect occurrence areas are marked, and a defect probability heatmap is generated as historical defect distribution information. Historical defect detection data includes data such as defect location coordinates, defect type labels, defect spectral characteristics, detection time, and batch. The defect probability heatmap is a visual distribution map based on historical defect distribution, using color intensity to represent the probability of defect occurrence.

[0029] Specifically, all historical defect detection data for this tire model are compiled, the actual location of each defect is marked, the frequency of defect occurrence in each area is counted, and a defect probability heatmap is generated through interpolation calculation as historical defect distribution information.

[0030] For example, the historical defect data of LT-21560R17-M02 over the past year is summarized: the defects are concentrated at the bottom of the main groove and the corners of the tread blocks; after statistics, a heat map is generated: the probability of defects at the bottom of the groove is 75%, the probability at the corners of the tread blocks is 60%, and the probability in the middle of the sidewall is 10%.

[0031] Furthermore, the geometric structure information, the optical prior information, and the historical defect distribution information are associated and stored with the tire model to construct a tire model prior structure map database. The tire model prior structure map refers to a standardized inspection benchmark map that integrates geometric structure information, optical prior information, and historical defect distribution information. Associated storage refers to binding the prior structure map with the corresponding tire model identifier to achieve a one-to-one correspondence. By integrating the geometric structure information, optical prior information, and historical defect distribution information to form the prior structure map for that model, and then associating it with the tire model identifier, a comprehensive database is constructed.

[0032] For example, the geometric structure information, optical prior information, and defect probability heat map corresponding to LT-21560R17-M02 are integrated into a dedicated prior structure map, which is then bound to the model identifier LT-21560R17-M02 and stored in the tire model prior structure map database.

[0033] Finally, based on the tire model identifier, the prior structure map corresponding to the tire model identifier is retrieved from the pre-built tire model prior structure map database. Using the tire model identifier loaded in S100 as the search keyword, a precise matching search is performed in the pre-built tire model prior structure map database to retrieve the prior structure map uniquely corresponding to that model.

[0034] For example, the tire model identifier LT-21560R17-M02 loaded in S100 is obtained. Using this identifier as a keyword, the database is searched to accurately match and retrieve the prior structure map corresponding to this model. The map contains complete geometric, optical, and defect history distribution information of the tire.

[0035] In this embodiment of the invention, three types of heterogeneous information—geometric structure, optical properties, and historical defect distribution—are organically integrated to form a multi-dimensional prior knowledge base specific to tire models, providing rich reference benchmarks for subsequent testing. Based on the precise retrieval of model identifiers, a one-to-one matching between the prior structure map and the tire to be tested is achieved, avoiding the problem of confusion in testing benchmarks for different tire models, improving the accuracy and efficiency of data retrieval, and effectively reducing the false detection rate and false negative rate of subsequent testing.

[0036] S300: Acquire multispectral images of the tire to be tested to obtain real-time multispectral images.

[0037] In this embodiment of the invention, multispectral images of the tire to be inspected are acquired to obtain real-time multispectral images. Tire surface defects are diverse, including surface cracks, rubber impurities, missing characters, and foreign objects at the bottom of grooves. A single spectral image can only reflect shallow surface texture and brightness information, offering limited identification of defects in hidden areas such as the bottom of grooves, subtle internal rubber defects, and low-contrast defects, thus failing to meet comprehensive inspection requirements. By acquiring multispectral images, the differentiated optical response characteristics of different bands to tire rubber materials and defect areas can be utilized to obtain multi-dimensional visual information. This provides rich and reliable real-time data support for subsequent image registration and regional defect detection, solving the problems of limited information and high false negative rates associated with single-spectral detection.

[0038] Specifically, an industrial multispectral camera is deployed at the tire inspection station. The camera's angle and focal length are adjusted to fully cover the tread and sidewall areas of the tire to be inspected. The acquisition bands of the multispectral camera are configured, and the simultaneous acquisition mode of visible light and near-infrared bands is enabled. A uniform ring light source is used to eliminate illumination shadows. After the tire is positioned, the multispectral camera is triggered to acquire a full-area image of the tire surface. The camera performs noise reduction and distortion correction preprocessing on the acquired multispectral raw images and integrates them to generate a real-time multispectral image in a standard format.

[0039] For example, for the tire to be inspected with model number LT-21560R17-M02, after completing the station positioning: an industrial multispectral camera equipped with visible light and near-infrared dual-bands is activated to acquire a full-area image of the tire; the visible light band image and the near-infrared band image of the tire are acquired simultaneously; after preprocessing, a real-time multispectral image containing dual-band data is generated as a real-time data source for subsequent registration and inspection.

[0040] In this embodiment of the invention, multi-band real-time images are acquired simultaneously by a multispectral camera, obtaining multi-dimensional optical information of the tire surface, which makes up for the lack of information in single-spectral images. This can effectively enhance the visual recognition of hidden areas and subtle defects, provide accurate real-time feature data for subsequent non-rigid registration, and provide adapted image data for differentiated detection strategies in different areas, thereby improving the comprehensiveness and accuracy of defect detection from the data source.

[0041] S400: Perform non-rigid registration between the prior structure map and the real-time multispectral image to obtain an aligned prior map that is aligned with the real-time multispectral image.

[0042] In this embodiment of the invention, the prior structure map and the real-time multispectral image are non-rigidly registered to obtain an aligned prior map that is aligned with the real-time multispectral image. Tires are flexible rubber products, and different tires of the same model may exhibit uneven stretching and twisting under varying inflation conditions, clamping angles, and surface tensions. Rigid registration cannot compensate for local deformations, leading to misalignment of character areas and groove boundaries in the standard map with the real-time image, resulting in incorrect region division. Therefore, non-rigid registration is necessary to accurately fit the standard prior structure map to the real-time multispectral image, eliminating global and local deformation errors and ensuring the accuracy of subsequent detection benchmarks.

[0043] Step S400 in the method provided in this embodiment of the invention includes: Several geometric feature points are extracted from the prior structure map and configured as an anchor point set, wherein the geometric feature points include character corner points, pattern block boundary points and groove turning points; Feature points corresponding to the anchor point set are identified from the real-time multispectral image to obtain a real-time feature point set; Based on the anchor point set and the real-time feature point set, the thin plate spline interpolation function is calculated to obtain the transformation mapping from the prior structure map to the real-time multispectral image; Each pixel coordinate in the prior structure map is converted into the corresponding coordinate in the real-time multispectral image through the transformation mapping to obtain the aligned prior map.

[0044] First, several geometric feature points are extracted from the prior structure map and configured as an anchor point set. These geometric feature points include character corner points, tread block boundary points, and groove inflection points. Character corner points refer to pixels at the intersections, endpoints, or corners of characters on the tire surface; they exhibit significant angular changes in the image, making them convenient as registration anchor points. Tread block boundary points refer to pixels on the boundary line between the raised tread area and the groove area, typically located at the inflection points of the tread block edge contour or at locations with significant curvature changes. Groove inflection points refer to locations where the direction of the tire surface grooves changes significantly, such as the start point, end point, bifurcation point, or apex of a bend in the groove.

[0045] First, the tread pattern outline and the coordinate set of the sidewall character region are obtained from the geometric structure information of the prior structure map. Based on the tread pattern outline, Canny edge detection combined with morphological thinning is used to extract the boundary lines between the tread blocks and the grooves. Significant turning points are automatically selected on the boundary lines according to the curvature changes to form a set of tread block boundary points. For areas with relatively gentle tread block boundaries, a uniform sampling method can be used to supplement a certain number of boundary points to ensure that the anchor points have sufficient coverage in spatial distribution.

[0046] Secondly, based on the coordinate set of the character region on the side of the tire, skeleton extraction or corner detection is performed on the character image in each character region, such as Harris corner detection or FAST corner detection. Pixels at the intersection of character strokes, endpoints, and significant corners are selected to form a set of character corners.

[0047] Next, the center line of the trench is extracted from the trench depth distribution map, and the locations where the trench direction changes significantly are identified, including the trench start point, end point, bifurcation point, and vertices with a large degree of curvature. The pixel coordinates corresponding to the locations where significant changes occur are extracted as a set of trench turning points.

[0048] Finally, the three types of feature points are merged, and redundant points that are too close in space are removed to form the final set of anchor points. Each anchor point records its pixel coordinates in the prior structure map coordinate system and is labeled with a feature type for subsequent matching.

[0049] For example, the prior structure map for the LT-21560R17-M02 tire model is as follows: character corner points are extracted: there are 8 stroke corners and intersections of the character 215 / 60R17 on the tire sidewall; tread block boundary points are extracted: there are 12 turning points of the edge contour of the main tread block; groove turning points are extracted: there are 6 bifurcations and bending vertices of the longitudinal main groove; the above 26 points are integrated to form the anchor point set for this model.

[0050] Secondly, feature points corresponding to the anchor point set are identified from the real-time multispectral image to obtain a real-time feature point set. The real-time feature point set refers to the actual feature points in the real-time multispectral image acquired by the S300 that correspond one-to-one with the prior anchor point set, reflecting the actual position and deformation state of the tire.

[0051] Specifically, for each anchor point in the anchor point set, the most suitable spectral band is selected for matching and positioning based on its feature type label. Character corner points typically have the highest contrast and clearest details in visible light images, so visible light images are preferred for character corner point matching; pattern block boundary points and groove inflection points have better edge contrast in near-infrared images, so near-infrared images can be used for matching. The matching method combines template matching with local feature descriptors: for each anchor point, a small-sized template image is cropped from its coordinates in the prior structure map, and normalized cross-correlation is calculated in the corresponding search area of ​​the real-time multispectral image, or accelerated robust features are used for feature point matching, and the location with the strongest response is used as the real-time feature point coordinates. For the matching results, a confidence evaluation mechanism can be introduced: if the matching response strength is lower than a set threshold, or the matching point does not match the anchor point type label, the point is considered a failed match and is discarded, ensuring that all anchor point pairs participating in the transformation calculation have high reliability. Finally, all successfully matched real-time feature points and their corresponding anchor points constitute control point pairs, forming a real-time feature point set.

[0052] For example, in the real-time multispectral image of LT-21560R17-M02: the actual coordinates of the corner points of the 8 characters corresponding to the characters on the sidewall are matched and identified; the actual coordinates of the boundary points of the 12 tread blocks are matched and identified; the actual coordinates of the 6 groove turning points are matched and identified; and a set of 26 real-time feature points corresponding one-to-one with the prior anchor point set is formed.

[0053] Next, based on the set of anchor points and the set of real-time feature points, a thin-plate spline interpolation function is calculated to obtain the transformation mapping from the prior structure map to the real-time multispectral image. The thin-plate spline interpolation function is a mathematical method for image registration and non-rigid deformation modeling. Its physical prototype can be analogized to bending an infinitely thin, elastic metal plate so that it precisely passes through several given anchor points while maintaining minimum bending energy at the remaining positions. In the tire detection scenario, this function establishes a mapping relationship between anchor points in the standard map and corresponding points in the real-time image, obtaining the transformation parameters by solving a system of linear equations. This transformation can simultaneously compensate for translation, rotation, scaling, and local elastic deformation, thereby accurately aligning the prior structure map to the actual captured tire image, achieving sub-pixel level registration accuracy.

[0054] Specifically, a nonlinear mapping relationship between the set of anchor points and the set of real-time feature points is established by constructing a thin-plate spline interpolation function. The specific construction process is as follows: Each anchor point and its corresponding real-time feature point form a control point pair. All control point pairs serve as constraints for the interpolation function. A transformation function that minimizes the bending energy of the elastic thin plate is solved. This function consists of two superimposed parts: a linear basis describing the overall affine transformation, plus a weighted combination of several radial basis functions to describe local nonlinear deformation. The radial basis functions are centered at each anchor point, and their weight coefficients are determined by solving a system of linear equations, ensuring that the function accurately passes through the corresponding real-time feature point coordinates at each anchor point. In non-control point regions, smooth interpolation is achieved by minimizing bending energy, thus obtaining a continuous transformation mapping that can map any coordinate in the prior structure map to the corresponding position in the real-time multispectral image.

[0055] For example, by substituting the 26 sets of anchor points and real-time feature points of LT-21560R17-M02 into the model, the thin plate spline interpolation function is calculated to generate the global and local deformation transformation mapping of the tire from the prior map to the real-time image, which can compensate for the coordinate transformation relationship of the tire rotating 3° and the slight deformation of the local tread pattern.

[0056] Finally, each pixel coordinate in the prior structure map is converted into its corresponding coordinate in the real-time multispectral image through the transformation mapping to obtain the aligned prior map. Using the transformation mapping, all pixel coordinates in the prior structure map are traversed and converted one by one into their corresponding coordinates in the real-time multispectral image, completing the non-rigid alignment of the entire image and generating the aligned prior map.

[0057] For example, each pixel of the LT-21560R17-M02 prior structure map is transformed and mapped to the corresponding position in the real-time image. The original standard geometric structure, optical prior, and defect history distribution information are all aligned with the actual shape of the tire, and finally, an aligned prior map is generated.

[0058] In this embodiment of the invention, by using non-rigid registration of thin plate splines based on geometric feature points, the translation, rotation, scaling, and local elastic deformation of the tire can be compensated simultaneously, achieving sub-pixel-level precise alignment between the prior structure map and the real-time multispectral image. This solves the problem of mismatch between the standard benchmark and the actual tire shape, allowing prior information such as geometric structure and defect distribution to accurately correspond to the actual tire area. This provides a precise alignment benchmark for subsequent dynamic region division and detection strategy allocation, fundamentally avoiding detection failures caused by misalignment.

[0059] S500: Based on the aligned prior map, the real-time multispectral image is dynamically divided into several regions of interest, and a corresponding detection strategy is assigned to each region of interest. The detection strategies include a high-precision defect detection strategy, an optical character recognition strategy, a rapid screening strategy, and a hidden area enhancement detection strategy.

[0060] In this embodiment of the invention, based on the aligned prior map, the real-time multispectral image is dynamically divided into several regions of interest (ROIs), and a corresponding detection strategy is assigned to each ROI. These detection strategies include a high-precision defect detection strategy, an optical character recognition strategy, a rapid screening strategy, and a hidden area enhancement detection strategy. The aligned prior map achieves precise alignment with the real-time multispectral image, but it contains information about the entire tire area. Different regions exhibit significant differences in defect risk, structural characteristics, and optical response: high-risk defect areas require precise detection, character recognition areas need to avoid misjudgments, low-risk areas need improved efficiency, and hidden areas need to overcome detection blind spots. Using a uniform detection strategy would result in low efficiency in high-precision areas, insufficient accuracy in high-efficiency areas, missed detections in blind spots, and misjudgments in character areas. Therefore, it is necessary to dynamically divide ROIs based on prior information and assign appropriate detection strategies to each region to achieve a balance between precise and efficient detection, and between blind spot coverage and misjudgment avoidance, ensuring the accuracy, efficiency, and comprehensiveness of the detection.

[0061] Step S500 in the method provided in this embodiment of the invention includes: Extract the defect probability heatmap from the defect history distribution information in the aligned prior map, divide the region with a defect probability greater than or equal to the first probability threshold into the first type of region of interest, and assign a high-precision defect detection strategy to the first type of region of interest. Extract the coordinates of the tire side character region from the geometric structure information in the aligned prior map, divide the tire side character region into a second type of region of interest, and assign an optical character recognition strategy to the second type of region of interest. Extract the tread pattern outline from the geometric structure information of the aligned prior map, divide the tread pattern protrusion area and the groove wall area into the third type of region of interest, and assign a fast screening strategy to the third type of region of interest. Extract the trench depth distribution map from the geometric structure information of the aligned prior map, divide the bottom region of the trench into a fourth type of region of interest, and assign a hidden region enhancement detection strategy to the fourth type of region of interest.

[0062] First, a defect probability heatmap is extracted from the defect history distribution information in the aligned prior map. Regions with defect probabilities greater than or equal to a first probability threshold are classified as first-class regions of interest, and high-precision defect detection strategies are assigned to the first-class regions of interest.

[0063] The high-precision defect detection strategy includes: Full-band multispectral image acquisition is enabled for the first type of region of interest; The acquired full-band multispectral images are input into a pre-trained defect recognition neural network model to obtain the defect location, defect type, and defect confidence level.

[0064] First, a defect probability heatmap is extracted from the historical defect distribution information of the aligned prior map. Regions with defect probabilities greater than or equal to a first probability threshold are classified as Regions of Interest (ROIs). A high-precision defect detection strategy is then assigned to these ROIs. The ROIs are high-incidence defect areas identified based on historical defect distribution information, corresponding to spatial locations with high defect probabilities in the prior map. These areas are the focus of defect detection and require the highest level of detection resources. The first probability threshold is a preset defect probability critical value used to distinguish between high-incidence and low-incidence defect areas, and can be dynamically adjusted according to tire model and manufacturing process. The high-precision defect detection strategy refers to a specialized detection method that utilizes full-band multispectral analysis and complex models to ensure that critical defects such as microcracks and early scratches are not missed in high-incidence defect areas.

[0065] Specifically, a defect probability heatmap is extracted from the defect history distribution information of the aligned prior map, and a preset first probability threshold is invoked. All regions of the heatmap are traversed, and regions with a defect probability ≥ the first probability threshold are marked and classified as Regions of Interest (ROIs). The first probability threshold is a pre-calibrated, dynamically defined critical parameter bound to the tire model; it is not a fixed value. Its preset process is entirely based on a comprehensive consideration of industrial measured data, historical production patterns, structural characteristics, and quality control requirements; for example, it may be set to 60%.

[0066] For example, for the aligned prior map of the LT-21560R17-M02 tire: extract the defect probability heatmap, call the first probability threshold of 60%, the probability of defects at the bottom of the groove is 75%, and the probability of defects at the edge of the tread block is 60%, and mark the bottom of the groove and the edge of the tread block as the first type of region of interest.

[0067] Secondly, full-band multispectral image acquisition is enabled for the first type of region of interest. Full-band multispectral image acquisition refers to using all preset bands of the multispectral camera to acquire complete optical information of the target area, providing sufficient data support for defect identification. The full-band acquisition mode of the industrial multispectral camera is enabled to acquire full-band images of the first type of region of interest. For example, full-band acquisition is enabled for the first type of region of interest, using visible light 400–700 nm + near-infrared 780–900 nm to obtain a complete multispectral image of the area.

[0068] Next, the acquired full-band multispectral images are input into a pre-trained defect recognition neural network model to obtain the defect location, defect type, and defect confidence level. The defect recognition neural network model is a deep learning model pre-trained with a large number of tire defect samples, capable of accurately identifying defect location, type, and confidence level. The acquired full-band images are then input into the pre-trained defect recognition neural network model, which outputs the defect location, defect type, and defect confidence level for the first type of region of interest.

[0069] The construction process of the defect recognition neural network model is as follows: First, the defect recognition neural network model is constructed using a lightweight convolutional neural network (CNN) to adapt to the multispectral defect detection scenario of tires, balancing detection accuracy with real-time requirements in industrial settings. The model's input layer is a 4-channel multispectral image with a uniform size of 224×224, comprising 3 convolutional hidden layers, 3 pooling layers, 2 fully connected hidden layers, and a three-branch output layer. The 3 convolutional hidden layers have 32, 64, and 128 3×3 convolutional kernels respectively, all using ReLU activation. Each convolutional layer is followed by a 2×2 max-pooling layer to compress feature dimensions. The 2 fully connected hidden layers have 256 and 128 neurons respectively, both using ReLU activation. The first fully connected layer is equipped with Dropout(0.5) to prevent overfitting. The output layer includes a defect type classification branch, a defect location regression branch, and a defect confidence prediction branch. The defect type classification branch outputs the defect category, the defect location regression branch outputs the defect coordinates, and the defect confidence prediction branch outputs the reliability of the current defect recognition result.

[0070] Secondly, the training dataset comes from three types of samples: real industrial defect multispectral images of various tire models collected on-site, expanded samples processed by rotation, flipping, spectral brightness adjustment, etc., and artificial defect standard samples prepared in the laboratory. The total dataset contains 100,000 labeled samples, which are divided into training set and validation set in an 8:2 ratio.

[0071] Furthermore, the model is trained using a combination of supervised learning and transfer learning. The model is initialized with pre-trained weights to accelerate training convergence. Adam is selected as the optimizer, with an initial learning rate of 0.001 and a batch size of 16. A batch iterative training mode is adopted, and the model performance is evaluated in real time using a validation set during training, and the training parameters are dynamically adjusted.

[0072] Finally, a multi-task weighted loss function is adopted to adapt to the dual tasks of defect type classification and defect location regression. By reasonably allocating the loss weights for classification and regression tasks, the model ensures that it simultaneously considers the accuracy of defect type identification and the precision of defect location. The convergence condition is that training stops when any of the following conditions are met: the improvement in validation set accuracy is less than 0.1% for 10 consecutive rounds; the decrease in total loss value is less than 0.001 for 15 consecutive rounds; or the preset maximum number of training rounds (50 rounds) is reached.

[0073] For example, the full-band multispectral image is converted to a uniform size of 224×224 and input into the trained defect recognition neural network model. The defect recognition neural network model outputs: Defect location: pixel coordinates (320, 450); Defect type: there is a small crack at the bottom of the trench; Defect confidence: 92%.

[0074] Secondly, the coordinates of the tire side character region in the geometric structure information are extracted from the aligned prior map, the tire side character region is divided into a second type of region of interest, and an optical character recognition strategy is assigned to the second type of region of interest.

[0075] The optical character recognition strategy includes: Acquire visible light images of the second type of region of interest; Perform optical character recognition on the visible light image to obtain character content recognition results; The character content recognition result is compared with the second type of standard characters of interest pre-stored in the prior structure map. If the comparison result is inconsistent, a character defect identifier is output.

[0076] First, the coordinates of the tire sidewall character regions are extracted from the geometric structure information of the aligned prior map. These tire sidewall character regions are then classified as second-type regions of interest (ROIs), and optical character recognition (OCR) strategies are assigned to them. The second-type ROI refers to the areas on the tire sidewall printed with characters such as specifications, brand, and model. The primary detection requirement for these areas is verifying the integrity and readability of the character content, rather than detecting surface defects. The OCR strategy refers to a specialized detection method that uses dedicated optical character recognition to replace general defect detection for character areas, avoiding misjudging character edges as cracks and reducing false alarm rates at the source. The coordinates of the tire sidewall character regions are extracted from the geometric structure information of the aligned prior map, and these regions are marked and classified as second-type ROIs based on their coordinate range.

[0077] For example, for the aligned prior map of the LT-21560R17-M02 tire model: extract the coordinates of the character region on the tire sidewall: top left corner (25,180), bottom right corner (180,320), and divide this rectangular region into the second type of region of interest.

[0078] Secondly, visible light images were acquired for the second type of region of interest. Characters have the highest recognizability in the visible light band, therefore only visible light images of the second type of region of interest were acquired.

[0079] Next, optical character recognition (OCR) is performed on the visible light image to obtain character content recognition results. Optical character recognition (OCR) is a technology that uses image recognition to convert characters in an image into editable text, suitable for the accurate recognition and comparison of characters on tire sidewalls. OCR recognition is performed on the visible light image to extract the character content and obtain the character recognition results.

[0080] For example, a visible light image of the area is acquired, and the character content 215 / 60R17-M03 is obtained through OCR recognition.

[0081] Finally, the character content recognition result is compared with the second type of standard characters of interest pre-stored in the prior structure map. If the comparison result is inconsistent, a character defect identifier is output. Standard characters refer to the standard sidewall characters corresponding to the tire model pre-stored in the prior structure map, such as specifications and brand identifiers, which serve as the benchmark for character comparison. The recognition result is compared character-by-character with the standard characters of that model pre-stored in the prior structure map. If inconsistencies exist, such as missing or misprinted characters, a character defect identifier is output.

[0082] For example, the identified character content 215 / 60R17-M03 is compared with the standard character 215 / 60R17-M02 pre-stored in the prior map. If the last character is found to be inconsistent, a character defect mark is output to avoid misjudging the character printing deviation as a crack.

[0083] Next, the tread pattern outline in the geometric structure information is extracted from the aligned prior map, and the tread pattern protrusion area and groove wall area are divided into third type of region of interest, and a fast screening strategy is assigned to the third type of region of interest.

[0084] The rapid screening strategy includes: Single-spectral visible light images were acquired for the third type of region of interest; Calculate the texture feature vector of the single-spectral visible light image; The texture feature vector is compared with the standard texture feature vector of the third type of region of interest pre-stored in the prior structure map to obtain the difference value. If the difference value is less than the preset difference threshold, the corresponding region is determined to be without defects. If the difference value is greater than or equal to the preset difference threshold, the high-precision defect detection strategy is re-executed for the third type of region of interest.

[0085] First, the tread pattern outline is extracted from the geometric structure information of the aligned prior map. The raised areas and groove wall areas of the tread pattern are classified as third-type regions of interest (ROIs), and a rapid screening strategy is assigned to these ROIs. The third-type ROIs refer to the raised areas and groove wall areas of the tread pattern. These areas have high structural strength and a relatively low probability of defect occurrence, making them suitable for lightweight rapid screening. The rapid screening strategy refers to a specialized detection method that uses only single-spectrum lightweight comparison for low-risk areas, saving computational resources, and triggering detailed inspection only when anomalies are detected, thereby improving overall detection efficiency.

[0086] For example, the tread pattern outline is extracted from the geometric structure information of the aligned prior map. Based on the outline boundaries, the raised areas and groove wall areas of the tread pattern are marked and classified as third-class regions of interest. For the aligned prior map of the LT-21560R17-M02 tire: the tread pattern outline is extracted, and the raised areas and groove wall areas of the four longitudinal main patterns are classified as third-class regions of interest.

[0087] Secondly, single-spectral visible light images are acquired for the third type of region of interest. Acquiring only single-spectral visible light images of the third type of region of interest eliminates the need for multi-band acquisition, simplifying the data processing flow and saving computing power.

[0088] Next, the texture feature vector of the single-spectrum visible light image is calculated. Texture features are extracted from the single-spectrum visible light image, and the texture feature vector of the visible light image is calculated, including contrast, roughness, and directionality.

[0089] For example, the single-spectral visible light image of the third type of region of interest is converted to grayscale, and the entire region is traversed using a 5×5 pixel local computation window: the contrast is calculated using the gray-level co-occurrence matrix method, resulting in a contrast of 82, indicating stable gray-level differences between adjacent pixels in the region; the roughness is calculated using the local gray-level variance method, resulting in a roughness of 11, indicating smooth and uniform surface in the region; and the directionality is calculated using the gradient direction statistics method, resulting in a directionality of 0.72, indicating regular and concentrated texture directions in the region. Combining these three feature values ​​in sequence yields the texture feature vector [82, 11, 0.72] for the region, which is then compared with the standard texture feature vector.

[0090] Finally, the texture feature vector is compared with the standard texture feature vector of the third type of region of interest pre-stored in the prior structure map to obtain the difference value. If the difference value is less than a preset difference threshold, the corresponding region is determined to be defect-free. If the difference value is greater than or equal to the preset difference threshold, the high-precision defect detection strategy is re-executed for the third type of region of interest. The preset difference threshold is a preset texture feature vector comparison threshold, such as 0.15, used to determine whether the region texture is abnormal. The smaller the difference value, the closer it is to the standard texture.

[0091] Specifically, the calculated texture feature vector is compared with the standard texture feature vector of the region pre-stored in the prior structure map, and the difference between the two is calculated, such as Euclidean distance or cosine distance. If the difference value is less than the preset difference threshold, the region is determined to be defect-free. If the difference value is greater than or equal to the preset difference threshold, the high-precision defect detection strategy is triggered to be re-executed in the region.

[0092] For example, the texture feature vector (contrast 82, roughness 11, directionality 0.72) is compared with the standard texture feature vector (contrast 85, roughness 12, directionality 0.75) pre-stored in the prior map, and the Euclidean distance is calculated. The difference value is, for example, 0.08. Since 0.08 < the preset difference threshold of 0.15, it is determined that there is no defect in the area, and no fine inspection is required, saving computing power and inspection time. If the difference value is 0.18, the high-precision defect detection strategy is triggered to re-inspect.

[0093] Finally, the trench depth distribution map is extracted from the geometric structure information of the aligned prior map, the bottom region of the trench is divided into a fourth type of region of interest, and a hidden region enhancement detection strategy is assigned to the fourth type of region of interest.

[0094] The concealed area enhanced detection strategy includes: The fourth type of region of interest is switched to polarized light illumination; Acquire near-infrared images of the fourth type of region of interest; The near-infrared band image is input into a pre-trained anomaly detection model to obtain the location and type of foreign object.

[0095] First, the groove depth distribution map is extracted from the geometric structure information of the aligned prior map. The bottom region of the groove is divided into a fourth type of region of interest (ROI), and a hidden area enhancement detection strategy is assigned to this fourth type of ROI. The fourth type of ROI refers to the bottom region of the groove. This type of region is located in the depression on the tire surface, where light is difficult to reach under normal lighting conditions, resulting in poor imaging quality. It is a hidden area that requires special lighting and imaging methods. The hidden area enhancement detection strategy refers to a specialized detection method that actively switches between polarized light and near-infrared imaging for areas such as the bottom of the groove where light is difficult to reach, overcoming shadow interference and achieving effective identification of foreign objects and bubbles.

[0096] For example, the trench depth distribution map is extracted from the geometric structure information of the aligned prior map, and the bottom area of ​​the trench with a depth ≥ 5 mm is marked and classified as the fourth type of region of interest.

[0097] Secondly, the fourth type of region of interest is switched to polarized light illumination. Polarized light illumination refers to filtering stray light and shadows using a polarized light filter to enhance the image contrast of hidden areas and clearly reveal the details at the bottom of the trench. The light source at the inspection station is controlled to switch to polarized light illumination mode to eliminate interference from shadows at the bottom of the trench.

[0098] Next, near-infrared images of the fourth type of region of interest are acquired. The near-infrared band (780–900 nm) can penetrate minor impurities on the tire surface, clearly revealing hidden defects such as foreign objects and bubbles at the bottom of the grooves. Acquiring near-infrared images of the fourth type of region of interest enhances the identification of hidden defects.

[0099] Finally, the near-infrared image is input into a pre-trained anomaly detection model to obtain the location and type of the foreign object. The anomaly detection model is a pre-trained deep learning model designed for anomalies in concealed areas, capable of accurately identifying the location and type of anomalies. The near-infrared image is input into the pre-trained anomaly detection model, which outputs the location and type of the foreign object.

[0100] The construction process of the anomaly detection model is as follows: First, the anomaly detection model is constructed using a lightweight convolutional neural network (CNN), adapted for anomaly detection scenarios in near-infrared images of hidden areas such as the bottom of tire grooves. It is specifically designed to identify hidden anomalies such as foreign objects and bubbles, balancing detection accuracy with real-time requirements in industrial settings. The model's input layer is a single-channel near-infrared image with a uniform size of 224×224. It consists of three convolutional hidden layers, three pooling layers, two fully connected hidden layers, and a dual-branch output layer. The three convolutional hidden layers use 32, 64, and 128 3×3 convolutional kernels respectively, all with ReLU activation. Each convolutional layer is followed by a 2×2 max-pooling layer to reduce feature dimensionality. The two fully connected hidden layers have 256 and 128 neurons respectively, using ReLU activation. Dropout (0.5) is added to the first fully connected layer to suppress overfitting. The output layer includes an anomaly type classification branch and a foreign object location regression branch. The classification branch outputs the anomaly category (e.g., foreign object, bubble), while the location regression branch outputs the coordinates of the anomaly.

[0101] Secondly, the training dataset consists of three types of samples: real near-infrared images of hidden tire anomalies collected on-site, augmented samples enhanced by rotation, flipping, and brightness adjustment, and artificial anomaly standard samples created in the laboratory. All samples have been manually labeled with the type and location of the foreign object. The total dataset size is 80,000 labeled images, divided into training and validation sets in an 8:2 ratio for model training and performance validation.

[0102] Furthermore, a combination of supervised learning and transfer learning was used for training. The model was initialized with pre-trained weights to accelerate the training convergence speed. The Adam optimizer was selected with an initial learning rate of 0.001 and a batch size of 16. Training was completed in batch iteratively, and the classification accuracy and localization accuracy of the model were evaluated in real time through the validation set during the training process, and the training parameters were dynamically adjusted.

[0103] Finally, a multi-task weighted loss function is adopted, taking into account both anomaly type classification and foreign object location regression tasks. By assigning reasonable weights to the classification and regression tasks, the model can stably output accurate foreign object types and locations simultaneously. Model training stops and converges when any of the following conditions are met: the validation set accuracy improves by less than 0.1% for 10 consecutive rounds; the total loss decreases by less than 0.001 for 15 consecutive rounds; or the preset maximum number of training rounds (50 rounds) is reached.

[0104] For example, when a near-infrared band image is input into a trained anomaly detection model, the model outputs: Location of foreign object: pixel coordinates (410, 520); Type of foreign object: There is a small rubber foreign object at the bottom of the trench.

[0105] In this embodiment of the invention, based on the geometric and defect history information of the aligned prior map, accurate dynamic division of four types of regions of interest is achieved. Differential detection strategies are assigned based on the characteristics of each region: a high-precision strategy ensures detection accuracy in high-defect areas, avoiding missed detections of critical defects; an OCR strategy avoids misjudgments in character areas, reducing the overall false alarm rate; a rapid screening strategy saves computational power in low-risk areas, improving detection efficiency; and a hidden area enhancement strategy overcomes detection blind spots, achieving comprehensive detection without dead angles. Overall, this achieves the goal of accurate, efficient, and comprehensive detection, providing a scientific and suitable execution basis for subsequent regional detection and final decision-making.

[0106] S600: Each region of interest is detected using the detection strategy corresponding to that region, and the detection result for each region of interest is obtained.

[0107] In this embodiment of the invention, a detection strategy corresponding to each region of interest is used to detect the corresponding region of interest, and the detection result of each region of interest is obtained. S500 has completed the dynamic division of four types of regions of interest and assigned a detection strategy adapted to the characteristics of each region. The four types of regions of interest have significant differences in defect risk, detection targets, and structural features: the first type of region of interest is a high-incidence area of ​​defects, requiring accurate identification of defect details; the second type of region of interest is a character area, requiring verification of character integrity; the third type of region of interest is a low-risk area, requiring efficient screening; and the fourth type of region of interest is a hidden area, requiring identification of hidden anomalies. For each region of interest, its matching detection strategy is executed one by one to accurately obtain the specific detection results for each region, providing data support for subsequent detection result aggregation and final quality judgment. The coordinate range and corresponding detection strategy of the four types of regions of interest divided by S500 are retrieved to clarify the one-to-one correspondence between each region and the detection strategy, ensuring no strategy mismatch or region omission.

[0108] For example, taking the tire to be inspected, model LT-21560R17-M02, as an example, a full-band multispectral image is acquired for the first type of region, and input into the defect recognition neural network model to obtain the detection results: defect location (320, 450), defect type is a micro crack, and defect confidence level is 92%; a visible light image is acquired for the second type of region, and the OCR recognition obtains the character "215 / 60R17-M02", which matches the standard character, and the detection result is that there is no defect in the character region; a single-spectral visible light image is acquired for the third type of region, and the texture feature vector is calculated and compared with the standard vector. The difference value is 0.08, which is less than the preset threshold of 0.15, and the detection result is that there is no defect in this region; for the fourth type of region, polarized light illumination is switched, a near-infrared image is acquired and input into the anomaly detection model to obtain the detection results: foreign object location (410, 520), and foreign object type is a small rubber foreign object; the detection results of the above four types of regions are classified and stored, and the corresponding region coordinates and types are labeled to complete the S600 detection operation.

[0109] In this embodiment of the invention, by strictly adhering to the characteristics of each region of interest and executing corresponding detection strategies one by one, the detection operation is made targeted and standardized. High-precision and concealment enhancement strategies ensure detection accuracy in high-defect and concealed areas, capturing minute defects and hidden foreign objects. Rapid screening strategies improve detection efficiency in low-risk areas, and optical character recognition strategies mitigate the risk of misjudgment in character areas. Ultimately, comprehensive, accurate, and traceable regional detection results are obtained, fully covering the entire tire inspection requirements, providing reliable and comprehensive data support for the S700's detection result aggregation and final quality judgment.

[0110] S700: Summarize the detection results of all regions of interest, make a final judgment on the summarized detection results based on the historical distribution information of defects in the aligned prior map, and output the tire quality judgment result.

[0111] In this embodiment of the invention, the detection results of all regions of interest are summarized, and a final judgment is made on the summarized detection results based on the historical distribution information of defects in the aligned prior map, outputting the tire quality judgment result. S600 has obtained the sub-regional detection results of four types of regions of interest, including suspected defects, normal areas, and character verification results. However, the sub-regional detection results have two shortcomings: first, some suspected defects may be false alarms; second, the historical defect patterns of each region of the tire are not combined to verify suspected defects, and direct summarization can easily lead to judgment bias, making it impossible to accurately determine the overall tire quality. Therefore, it is necessary to first summarize all sub-regional detection results, and then combine them with the historical defect distribution information of the aligned prior map to scientifically verify and judge suspected defects, filter false alarms, confirm real defects, and finally output accurate and complete tire quality judgment results, ensuring the reliability of tire quality detection.

[0112] Step S700 in the method provided in this embodiment of the invention includes: Load the historical distribution information of defects in the aligned prior map to obtain the historical probability of defects occurring in each region of interest; If a suspected defect is detected in a region of interest, and the historical probability of the corresponding defect in the region of interest is greater than or equal to the second probability threshold, then the suspected defect is determined to be a real defect, and the defect location and defect type are marked; wherein, the output of the high-precision defect detection strategy and the hidden region enhancement detection strategy are the suspected defects; If a suspected defect is detected in a region of interest, and the historical probability of the corresponding defect in the region of interest is less than the second probability threshold, a second review is triggered. If the result of the second review is consistent with the result of the first detection, the suspected defect is determined to be a real defect. If the result of the second review is inconsistent with the result of the first detection, the suspected defect is determined to be a false alarm and filtered out. Summarize all information identified as genuine defects and output tire quality assessment results, which include non-conformance markings and defect distribution maps.

[0113] First, the historical defect distribution information in the aligned prior map is loaded to obtain the historical defect occurrence probability for each region of interest. The historical defect distribution information refers to the defect distribution data for this tire model stored in the aligned prior map within a preset historical time period, including a defect probability heatmap. The historical defect occurrence probability is the ratio of the defect statistical frequency in a region of interest to the total area of ​​that region, used to characterize the inherent risk of defects occurring in that region, with a value ranging from 0 to 100%.

[0114] Specifically, historical defect distribution information is extracted from the aligned prior map, which includes a defect probability heatmap. Based on the defect probability heatmap, the statistical frequency of actual defect occurrences in each region of interest within a preset historical time period is determined. The ratio of the statistical frequency to the total area of ​​the region is calculated to obtain the historical defect occurrence probability of the region of interest. The historical defect occurrence probability is associated with and stored in relation to the region of interest, and used for confidence weighting of suspected defects detected in the region during the final decision.

[0115] For example, for the tire to be inspected with model number LT-21560R17-M02: extract the defect probability heatmap from the aligned prior map; count the defect frequency of each region of interest within a preset historical time period: 80 times for the first type region, 0 times for the second type region, 12 times for the third type region, and 75 times for the fourth type region; calculate the historical defect occurrence probability: 75% for the first type region, 10% for the second type region, 10% for the third type region, and 75% for the fourth type region, with the second probability threshold set to 50%; associate and store the probability of each region with the corresponding region of interest for subsequent decision-making.

[0116] Secondly, if a suspected defect is detected in a region of interest, and the historical defect occurrence probability of the corresponding region of interest is greater than or equal to the second probability threshold, then the suspected defect is determined to be a real defect, and the defect location and defect type are marked. The output of the high-precision defect detection strategy and the hidden area enhancement detection strategy constitutes the suspected defect. A suspected defect is an anomaly information to be verified, containing the location and type of the defect / foreign object, but not yet confirmed for authenticity. The second probability threshold refers to a preset historical defect occurrence probability threshold, for example, 50%, used to distinguish between high-defect-risk areas and low-defect-risk areas, and can be dynamically adjusted according to the tire model.

[0117] Specifically, retrieve the regional detection results output by S600 and filter out suspected defects; retrieve the historical defect occurrence probability of each region and determine whether the probability of the region where the suspected defect is located is ≥ the second probability threshold; if the historical defect occurrence probability of the region is ≥ the second probability threshold, directly determine the suspected defect as a real defect and accurately mark the location and type of the defect / foreign object; record the determination result and proceed to the subsequent summary steps.

[0118] For example, for suspected defects of the LT-21560R17-M02 tire: screen suspected defects in the S600 inspection results: microcracks in the first category area and small rubber foreign objects in the fourth category area; set a second probability threshold of 50% and determine that the historical defect occurrence probability of both categories is ≥50%; directly determine the above two suspected defects as real defects and mark them as: microcracks (320, 450) and small rubber foreign objects (410, 520); record the determination result and proceed to the next step of summarization.

[0119] Secondly, if a suspected defect is detected in a region of interest (ROI), and the historical defect occurrence probability of the corresponding ROI is less than the second probability threshold, a second verification is triggered. If the second verification result is consistent with the initial detection result, the suspected defect is determined to be a real defect. If the second verification result is inconsistent with the initial detection result, the suspected defect is determined to be a false alarm and filtered out. A low-probability region refers to an ROI where the historical defect occurrence probability is less than the second probability threshold. The second verification involves re-executing the detection strategy for the corresponding region for suspected defects appearing in a low-probability region to verify the authenticity of the initial detection result and avoid false alarms.

[0120] Specifically, in the S600 detection results, suspected defects located in the area where the historical defect occurrence probability is less than the second probability threshold are screened out; for the area where the suspected defect is located, the corresponding detection strategy is re-executed to obtain the second verification result; compare the initial detection result with the second verification result: if both are the same and are suspected defects, then the suspected defect is determined to be a real defect; if the two are different, then it is determined to be a false alarm and is filtered out; record the determination result and proceed to the subsequent summary steps.

[0121] For example, assuming the historical defect occurrence probability of the third category area of ​​the LT-21560R17-M02 tire is 10% < 50%, and the texture feature vector difference value is 0.16 ≥ 0.15 during the initial S600 detection, a suspected defect is output after triggering high-precision detection; the suspected defect is determined to be located in a low-probability area, and a second review is initiated; the rapid screening strategy is re-executed on the third category area, a single-spectrum visible light image is acquired, the texture feature vector is calculated, and after comparison, the difference value is 0.14 < 0.15, and the second review result is no defect; the inconsistency between the initial and second results indicates that the suspected defect is a false alarm and is filtered out; the false alarm result is recorded and not included in the summary of real defects.

[0122] Finally, all information identified as genuine defects is summarized, and the tire quality assessment result is output. This result includes a non-compliance label and a defect distribution map. All identified genuine defects are summarized, and their location and type information are integrated to form a complete list of genuine defects. If the list is not empty, a non-compliance label is used; if the list is empty, the tire is deemed to be of acceptable quality. Based on real-time multispectral images, the location and type of all genuine defects are marked on the image to generate a defect distribution map, which is then integrated with the non-compliance label. A complete tire quality assessment result is output, including the non-compliance label, defect distribution map, and genuine defect list, for subsequent quality control and processing.

[0123] For example, for the LT-21560R17-M02 tire: Summarize the actual defects: microcracks (locations 320, 450) and small rubber foreign objects (locations 410, 520), forming a list of actual defects; because of the existence of actual defects, mark them with a non-conforming label; on the real-time multispectral image of the tire, mark the location and type of the above two defects, generating a defect distribution map; output the tire quality judgment result: non-conforming label + defect distribution map + list of actual defects, clearly indicating that the tire has microcracks and rubber foreign objects, and does not meet the quality standards.

[0124] In this embodiment of the invention, by summarizing the detection results by region and combining them with historical defect distribution information, suspected defects are verified and judged in a stratified manner. This effectively solves the problem of false alarms that may occur in regional detection and improves the accuracy of defect judgment. The stratified strategy of directly confirming high-probability areas and conducting secondary verification for low-probability areas avoids missing real defects and filters out false anomalies, ensuring the scientific nature of the judgment. The final tire quality judgment result presents the tire's defect situation completely and clearly, providing accurate and reliable evidence for tire quality control and handling of non-conforming products. It also enables defect traceability, completing the entire tire surface defect detection process in a closed loop, thus improving the reliability and practicality of the detection results.

[0125] Through the specific implementation methods described above, the embodiments of the present invention achieve the following technical effects: This invention provides a tire surface defect detection method and system based on multispectral image fusion. It utilizes a dedicated prior structure map precisely based on the tire model to provide a unified and reliable geometric, optical, and defect history benchmark for detection. Through multispectral image acquisition and non-rigid registration, it effectively eliminates detection deviations caused by tire deformation and pose shifts, achieving sub-pixel-level precise alignment between the prior benchmark and the actual tire. Based on regional characteristics, it dynamically divides regions of interest and assigns differentiated strategies such as high-precision detection, optical character recognition, rapid screening, and hidden area enhancement. This ensures detection accuracy in high-defect areas and hidden areas while improving detection efficiency in low-risk areas and preventing misjudgments of character regions from the outset. Finally, it combines historical defect distribution information to perform graded verification and secondary review of suspected defects, effectively filtering false alarms, identifying true defects, and reducing the false negative and false positive rates. This invention achieves full coverage, accurate identification, efficient operation, and stable reliability in tire surface defect detection, improving the intelligence and standardization of tire quality inspection.

[0126] Example 2, as Figure 2 As shown, this invention provides a tire surface defect detection system based on multispectral image fusion, the system comprising: The identifier acquisition module 11 is used to load the tire model identifier of the tire to be detected; The map retrieval module 12 is used to retrieve the prior structure map corresponding to the tire model identifier from a pre-built tire model prior structure map database based on the tire model identifier. The prior structure map includes geometric structure information, optical prior information, and defect history distribution information. Image acquisition module 13 is used to acquire multispectral images of the tire to be detected and obtain real-time multispectral images; The map alignment module 14 is used to perform non-rigid registration between the prior structure map and the real-time multispectral image to obtain an aligned prior map that is aligned with the real-time multispectral image. The region segmentation module 15 is used to dynamically divide the real-time multispectral image into several regions of interest based on the aligned prior map, and assign a corresponding detection strategy to each region of interest. The detection strategies include a high-precision defect detection strategy, an optical character recognition strategy, a rapid screening strategy, and a hidden area enhancement detection strategy. The region detection module 16 is used to detect the corresponding region of interest by using the detection strategy corresponding to each region of interest, and to obtain the detection result of each region of interest. The result output module 17 is used to summarize the detection results of all regions of interest, make a final judgment on the summarized detection results based on the defect history distribution information in the aligned prior map, and output the tire quality judgment result.

[0127] In one embodiment, the map invocation module 12 is further configured to: The construction of a priori structure map database for tire models includes: For each tire model, the geometric structure information of the tire model is extracted, wherein the geometric structure information includes the tread pattern outline, the coordinates of the sidewall character area, the groove depth distribution map, and the tread rubber thickness distribution map; Images of standard tire models of the specified tire type were collected under multiple spectra, and typical reflectance ranges and texture feature vectors of each region under different spectra were extracted as optical prior information. Summarize the historical defect detection data of the tire models, mark the actual areas where defects occurred, and generate a defect probability heatmap as historical defect distribution information; The geometric structure information, the optical prior information, and the defect history distribution information are associated and stored with the tire model to construct a tire model prior structure map database.

[0128] In one embodiment, the map alignment module 14 is further configured to: Several geometric feature points are extracted from the prior structure map and configured as an anchor point set, wherein the geometric feature points include character corner points, pattern block boundary points and groove turning points; Feature points corresponding to the anchor point set are identified from the real-time multispectral image to obtain a real-time feature point set; Based on the anchor point set and the real-time feature point set, the thin plate spline interpolation function is calculated to obtain the transformation mapping from the prior structure map to the real-time multispectral image; Each pixel coordinate in the prior structure map is converted into the corresponding coordinate in the real-time multispectral image through the transformation mapping to obtain the aligned prior map.

[0129] In one embodiment, the region division module 15 is further configured to: Extract the defect probability heatmap from the defect history distribution information in the aligned prior map, divide the region with a defect probability greater than or equal to the first probability threshold into the first type of region of interest, and assign a high-precision defect detection strategy to the first type of region of interest. Extract the coordinates of the tire side character region from the geometric structure information in the aligned prior map, divide the tire side character region into a second type of region of interest, and assign an optical character recognition strategy to the second type of region of interest. Extract the tread pattern outline from the geometric structure information of the aligned prior map, divide the tread pattern protrusion area and the groove wall area into the third type of region of interest, and assign a fast screening strategy to the third type of region of interest. Extract the trench depth distribution map from the geometric structure information of the aligned prior map, divide the bottom region of the trench into a fourth type of region of interest, and assign a hidden region enhancement detection strategy to the fourth type of region of interest.

[0130] The high-precision defect detection strategy includes: Full-band multispectral image acquisition is enabled for the first type of region of interest; The acquired full-band multispectral images are input into a pre-trained defect recognition neural network model to obtain the defect location, defect type, and defect confidence level.

[0131] The optical character recognition strategy includes: Acquire visible light images of the second type of region of interest; Perform optical character recognition on the visible light image to obtain character content recognition results; The character content recognition result is compared with the second type of standard characters of interest pre-stored in the prior structure map. If the comparison result is inconsistent, a character defect identifier is output.

[0132] The rapid screening strategy includes: Single-spectral visible light images were acquired for the third type of region of interest; Calculate the texture feature vector of the single-spectral visible light image; The texture feature vector is compared with the standard texture feature vector of the third type of region of interest pre-stored in the prior structure map to obtain the difference value. If the difference value is less than the preset difference threshold, the corresponding region is determined to be without defects. If the difference value is greater than or equal to the preset difference threshold, the high-precision defect detection strategy is re-executed for the third type of region of interest.

[0133] The concealed area enhanced detection strategy includes: The fourth type of region of interest is switched to polarized light illumination; Acquire near-infrared images of the fourth type of region of interest; The near-infrared band image is input into a pre-trained anomaly detection model to obtain the location and type of foreign object.

[0134] In one embodiment, the result output module 17 is further configured to: Load the historical distribution information of defects in the aligned prior map to obtain the historical probability of defects occurring in each region of interest; If a suspected defect is detected in a region of interest, and the historical probability of the corresponding defect in the region of interest is greater than or equal to the second probability threshold, then the suspected defect is determined to be a real defect, and the defect location and defect type are marked; wherein, the output of the high-precision defect detection strategy and the hidden region enhancement detection strategy are the suspected defects; If a suspected defect is detected in a region of interest, and the historical probability of the corresponding defect in the region of interest is less than the second probability threshold, a second review is triggered. If the result of the second review is consistent with the result of the first detection, the suspected defect is determined to be a real defect. If the result of the second review is inconsistent with the result of the first detection, the suspected defect is determined to be a false alarm and filtered out. Summarize all information identified as genuine defects and output tire quality assessment results, which include non-conformance markings and defect distribution maps.

[0135] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A tire surface defect detection method based on multispectral image fusion, characterized in that, include: Load the tire model identification of the tire to be tested; Based on the tire model identifier, the prior structure map corresponding to the tire model identifier is retrieved from the pre-built tire model prior structure map database, wherein the prior structure map includes geometric structure information, optical prior information and defect history distribution information; Acquire multispectral images of the tire to be tested to obtain real-time multispectral images; The prior structure map is non-rigidly registered with the real-time multispectral image to obtain an aligned prior map that is aligned with the real-time multispectral image. Based on the aligned prior map, the real-time multispectral image is dynamically divided into several regions of interest, and a corresponding detection strategy is assigned to each region of interest. The detection strategies include a high-precision defect detection strategy, an optical character recognition strategy, a rapid screening strategy, and a hidden area enhancement detection strategy. Each region of interest is detected using the detection strategy corresponding to that region, and the detection results for each region of interest are obtained. The detection results of all regions of interest are summarized, and a final judgment is made on the summarized detection results based on the historical distribution information of defects in the aligned prior map, and the tire quality judgment result is output.

2. The tire surface defect detection method based on multispectral image fusion as described in claim 1, characterized in that, Construct a priori structure map database for tire models, including: For each tire model, the geometric structure information of the tire model is extracted, wherein the geometric structure information includes the tread pattern outline, the coordinates of the sidewall character area, the groove depth distribution map, and the tread rubber thickness distribution map; Images of standard tire models of the specified tire type were collected under multiple spectra, and typical reflectance ranges and texture feature vectors of each region under different spectra were extracted as optical prior information. Summarize the historical defect detection data of the tire models, mark the actual areas where defects occurred, and generate a defect probability heatmap as historical defect distribution information; The geometric structure information, the optical prior information, and the defect history distribution information are associated and stored with the tire model to construct a tire model prior structure map database.

3. The tire surface defect detection method based on multispectral image fusion as described in claim 1, characterized in that, The prior structure map is non-rigidly registered with the real-time multispectral image to obtain an aligned prior map that is aligned with the real-time multispectral image, including: Several geometric feature points are extracted from the prior structure map and configured as an anchor point set, wherein the geometric feature points include character corner points, pattern block boundary points and groove turning points; Feature points corresponding to the anchor point set are identified from the real-time multispectral image to obtain a real-time feature point set; Based on the anchor point set and the real-time feature point set, the thin plate spline interpolation function is calculated to obtain the transformation mapping from the prior structure map to the real-time multispectral image; Each pixel coordinate in the prior structure map is converted into the corresponding coordinate in the real-time multispectral image through the transformation mapping to obtain the aligned prior map.

4. The tire surface defect detection method based on multispectral image fusion as described in claim 1, characterized in that, Based on the aligned prior map, the real-time multispectral image is dynamically divided into several regions of interest, and a corresponding detection strategy is assigned to each region of interest, including: Extract the defect probability heatmap from the defect history distribution information in the aligned prior map, divide the region with a defect probability greater than or equal to the first probability threshold into the first type of region of interest, and assign a high-precision defect detection strategy to the first type of region of interest. Extract the coordinates of the tire side character region from the geometric structure information in the aligned prior map, divide the tire side character region into a second type of region of interest, and assign an optical character recognition strategy to the second type of region of interest. Extract the tread pattern outline from the geometric structure information of the aligned prior map, divide the tread pattern protrusion area and the groove wall area into the third type of region of interest, and assign a fast screening strategy to the third type of region of interest. Extract the trench depth distribution map from the geometric structure information of the aligned prior map, divide the bottom region of the trench into a fourth type of region of interest, and assign a hidden region enhancement detection strategy to the fourth type of region of interest.

5. The tire surface defect detection method based on multispectral image fusion as described in claim 4, characterized in that, The high-precision defect detection strategy includes: Full-band multispectral image acquisition is enabled for the first type of region of interest; The acquired full-band multispectral images are input into a pre-trained defect recognition neural network model to obtain the defect location, defect type, and defect confidence level.

6. The tire surface defect detection method based on multispectral image fusion as described in claim 4, characterized in that, The optical character recognition strategy includes: Acquire visible light images of the second type of region of interest; Perform optical character recognition on the visible light image to obtain character content recognition results; The character content recognition result is compared with the second type of standard characters of interest pre-stored in the prior structure map. If the comparison result is inconsistent, a character defect identifier is output.

7. The tire surface defect detection method based on multispectral image fusion as described in claim 4, characterized in that, The rapid screening strategy includes: Single-spectral visible light images were acquired for the third type of region of interest; Calculate the texture feature vector of the single-spectral visible light image; The texture feature vector is compared with the standard texture feature vector of the third type of region of interest pre-stored in the prior structure map to obtain the difference value. If the difference value is less than the preset difference threshold, the corresponding region is determined to be without defects. If the difference value is greater than or equal to the preset difference threshold, the high-precision defect detection strategy is re-executed for the third type of region of interest.

8. The tire surface defect detection method based on multispectral image fusion as described in claim 4, characterized in that, The concealed area enhanced detection strategy includes: The fourth type of region of interest is switched to polarized light illumination; Acquire near-infrared images of the fourth type of region of interest; The near-infrared band image is input into a pre-trained anomaly detection model to obtain the location and type of foreign object.

9. The tire surface defect detection method based on multispectral image fusion as described in claim 1, characterized in that, The detection results of all regions of interest are summarized, and a final judgment is made on the summarized detection results based on the defect history distribution information in the aligned prior map, outputting the tire quality assessment result, including: Load the historical distribution information of defects in the aligned prior map to obtain the historical probability of defects occurring in each region of interest; If a suspected defect is detected in a region of interest, and the historical probability of the corresponding defect in the region of interest is greater than or equal to the second probability threshold, then the suspected defect is determined to be a real defect, and the defect location and defect type are marked; wherein, the output of the high-precision defect detection strategy and the hidden region enhancement detection strategy are the suspected defects; If a suspected defect is detected in a region of interest, and the historical probability of the corresponding defect in the region of interest is less than the second probability threshold, a second review is triggered. If the result of the second review is consistent with the result of the first detection, the suspected defect is determined to be a real defect. If the result of the second review is inconsistent with the result of the first detection, the suspected defect is determined to be a false alarm and filtered out. Summarize all information identified as genuine defects and output tire quality assessment results, which include non-conformance markings and defect distribution maps.

10. A tire surface defect detection system based on multispectral image fusion, characterized in that, The method for detecting tire surface defects based on multispectral image fusion according to any one of claims 1-9 includes: The identifier acquisition module is used to load the tire model identifier of the tire to be detected; The map retrieval module is used to retrieve the prior structure map corresponding to the tire model identifier from a pre-built tire model prior structure map database based on the tire model identifier. The prior structure map includes geometric structure information, optical prior information, and defect history distribution information. The image acquisition module is used to acquire multispectral images of the tire to be inspected, and obtain real-time multispectral images; The map alignment module is used to perform non-rigid registration between the prior structure map and the real-time multispectral image to obtain an aligned prior map that is aligned with the real-time multispectral image. The region segmentation module is used to dynamically divide the real-time multispectral image into several regions of interest based on the aligned prior map, and assign a corresponding detection strategy to each region of interest. The detection strategies include a high-precision defect detection strategy, an optical character recognition strategy, a rapid screening strategy, and a hidden area enhancement detection strategy. The region detection module is used to detect the corresponding region of interest by using the detection strategy corresponding to each region of interest, and to obtain the detection result of each region of interest; The result output module is used to summarize the detection results of all regions of interest, make a final judgment on the summarized detection results based on the defect history distribution information in the aligned prior map, and output the tire quality judgment result.

Images